Haar wavelets, fluctuations and structure functions ... · Haar wavelets, fluctuations and...
Transcript of Haar wavelets, fluctuations and structure functions ... · Haar wavelets, fluctuations and...
1
Haar wavelets, fluctuations and structure functions:
convenient choices for geophysics
S. Lovejoy,
Physics, McGill University, 3600 University st.,
Montreal, Que. H3A 2T8, Canada
D. Schertzer,
LEESU, École des Ponts ParisTech, Université Paris-Est, 6-8 Av. B. Pascal,
77455 Marne-la-Vallée Cedex 2, France
2
Abstract:
Geophysical processes are typically variable over huge ranges of space-time
scales. This has lead to the development of many techniques for decomposing series and
fields into fluctuations Δv at well defined scales. Classically, one defines fluctuations as
differences: (Δv)diff = v(x+Δx)-v(x) and this is adequate for many applications (Δx is the
“lag”). However, if over a range one has scalingΔv ∝ ΔxH , these difference fluctuations
are only adequate when 0<H<1, hence the need for other types of fluctuations. In
particular, atmospheric processes in the range ≈ 10 days to 10 years generally have -1 <
H < 0 so that a definition valid over the range -1< H <1 would be very useful for
atmospheric applications.
The general framework for defining fluctuations is wavelets. However the
generality of wavelets often leads to fairly arbitrary choices of “mother wavelet” and the
resulting wavelet coefficients may be difficult to interpret. In this paper we argue that a
good choice is provided by the (historically) first wavelet, the Haar wavelet (Haar,
1910) which is easy to interpret and - if needed - to generalize, yet has rarely been used
in geophysics. It is also easy to implement numerically: the Haar fluctuation (Δv)Haar at
lag Δx is simply proportional to the difference of the mean from x to x+Δx/2 and from
x+Δx/2 to x+Δx.
Using numerical multifractal simulations, we show that it is quite accurate, and
we compare and contrast it with another similar technique, Detrended Fluctuation
Analysis. We find that for estimating scaling exponents, the two methods are very
similar, yet Haar based methods have the advantage of being numerically faster,
theoretically simpler and physically easier to interpret.
1. Introduction:
In many branches of science - and especially in the geosciences - one commonly
decomposes space-time signals into components with well-defined space-time scales. A
3
widely used method is Fourier analysis which - when combined with statistics - can be
used to quantify the contribution to the variance of structures of a given frequency and/or
wavenumber to the total variance of the process. Often however, real space analyses are
preferable, not only because they are simpler to perform, but more importantly - because
they are simpler to interpret. The latter are based on various definitions of fluctuations at
a given scale and location; the simplest being the differences:
Δv Δx( )( )diff ≡ δΔxv ; δΔxv = v x + Δx( )− v x( ) (1)
where the position is x, and the scale (“lag”) Δx and δ is the difference operator (the
parentheses with the “diff” subscript is only used when it is important to distinguish the
difference fluctuation from others). For simplicity, we consider a scalar function (v) of a
scalar position and/or of time. We consider processes whose statistics are translationally
invariant (statistically homogeneous in space, stationary in time). The resulting
fluctuations can then be statistically characterized, the standard method being via the
moments <Δvq>. When q = 2, these are the classical structure functions, when q ≠2, they
are the “generalized” counterparts.
The space-time variability of natural systems, can often be broken up into various
“scaling ranges” over which the fluctuations vary in a power law manner with respect to
scale. Over these ranges, the fluctuations follow:
Δv = ϕΔvΔxH (2)
where ϕΔx is a random field (or series, or transects thereof at resolution Δx). Below for
simplicity, we consider spatial transects with lag Δx, but the results are identical for
temporal lags Δt. In fluid systems such as the atmosphere, ϕ is a turbulent flux. For
example, the Kolmogorov law corresponds to taking v as a velocity component and
ϕ = ε1/3 with ε as the energy flux. The classical turbulence laws assumed fairly uniform
fluxes, typically that ϕ is a quasi Gaussian process.
Taking the qth power of eq. 2 and ensemble averaging, we see that the statistical
characterization of the fluctuations in terms of generalized structure functions is:
4
Δv Δx( )q = ϕΔxq ΔxqH ≈ Δxξ q( ); ϕΔx
q = LΔx
⎛⎝⎜
⎞⎠⎟K q( ); ξ q( ) = qH − K q( )
(3) where ξ(q) is the structure function, K(q) is the moment scaling function that
characterizes the flux and L is that outer scale of the scaling regime (e.g. of a cascade
process). In the classical quasi-Gaussian case, K(q) = 0 so that ξ(q) is linear. More
generally, if the field is intermittent - for example if it is the result of a multifractal
process - then the exponent K(q) is generally nonzero and convex and characterizes the
intermittency. Since the ensemble mean of the flux spatially averaged at any scale is the
same (i.e. < φΔx > is constant, independent of Δx), we have K(1) = 0 and ξ(1) = H. The
physical significance of H is thus that it determines the rate at which mean fluctuations
grow (H>0) or decrease (H<0) with scale Δx.
There is a simple and useful expression for the exponent β of the power spectrum
E(k)≈ k-β. It is β =1+ξ(2) = 1+2H-K(2) which is a consequence of the fact that the
spectrum is a second order moment (the Fourier transform of the autocorrelation
function). K(2) is therefore often regarded as the spectral “intermittency correction”.
Note however that even if K(2) is not so large (it is typically in the range 0.1-0.2 for
turbulent fields), that the intermittency is nevertheless important: first because K(q)
typically grows rapidly with q so that the high order moments (characterizing the
extremes) can be very large, secondly because it is an exponent so that even when it is
small, whenever the range of scales over which it acts is large enough, its effect can be
large; in geo-systems the scale range can be 109 or larger.
Over the last thirty years, structure functions based on differences and with
concave rather than linear ξ(q) have been successfully applied to many geophysical
processes; this is especially true for atmospheric processes at time scales shorter than ≈
10 days which usually have 0 > H > 1. However, it turns out that the use of differences
to define the fluctuations is overly (and needlessly) restrictive. The problem becomes
evident when we consider that the mean difference cannot decrease with increasing Δx:
for processes with H<0, the differences simply converge to a spurious constant depending
5
on the highest frequencies available in the sample. Similarly differences cannot grow
faster than Δx – they “saturate” at a large value - independent of Δx which depends on the
lowest frequencies present in the sample, hence they cannot be used when H >1. Overall,
whenever H is outside the range 0>H>1, the exponent ξ(q) is no longer correctly
estimated. The problem is that we need a definition of fluctuations such that Δv(Δx) is
dominated by wavenumbers in the neighbourhood of Δx-1. Appendix A gives more
details about the relation between real and Fourier space and the ranges of H relevant for
various types of fluctuation and fig. 1 shows the shape of typical sample series as H
varies.
The need to define fluctuations more flexibly motivated the development of
wavelets (e.g. (Holschneider, 1995), (Torrence and Compo, 1998)), the classical
difference fluctuation being only a special case, the “poor man’s wavelet”. To change the
range of H over which fluctuations are usefully defined one must change the shape of the
defining wavelet. In the usual wavelet framework, this is done by modifying the “mother
wavelet” directly e.g. by choosing various derivatives of the Gaussian (e.g. the second
derivative “Mexican hat”). Following this, the fluctuations are calculated as convolutions
with Fast Fourier (or equivalent) numerical techniques.
A problem with this usual wavelet implementation is that not only are the
convolutions numerically cumbersome, but the physical interpretation of the fluctuations
is largely lost. In contrast, when 0<H<1, the difference structure function gives direct
information on the typical difference (q = 1) and typical variations around this difference
(q = 2) and even typical skewness (q = 3) or typical Kurtosis (q = 4) or - if the probability
tail is algebraic – of the divergence of high order moments of differences. Similarly,
when -1 < H < 0 one can define the “tendency structure function” (below) which directly
quantifies the fluctuation’s deviation from zero and whose exponent characterizes the rate
at which the deviations decrease when we average to larger and larger scales. These poor
man’s and tendency fluctuations are also very easy to directly estimate from series with
uniformly spaced data and - with straightforward modifications - to irregularly spaced
data.
6
In (Lovejoy and Schertzer, 2012a), (Lovejoy and Schertzer, 2012b) it is
argued that with only a few exceptions, for atmospheric processes in the “weather”
regime (Δt <≈10 days), that 0 < H <1 so that fluctuations increase with scale and the
classical use of differences is adequate. However, in the low frequency weather regime
(10 days ≈< Δt <≈ 10- 30 years), on the contrary, quite generally (in time, but not in
space!) we find -1< H < 0 so that fluctuations decrease with scale (although for larger Δt -
the “climate” regime – it is again in the range 0 > H > 1). In order to cover the whole
range (to even millions of years) one must thus use a definition of fluctuations which is
valid over the range 1> H > -1. This range is also appropriate for many solid earth
applications, including topography, geogravity and geomagnetism, for a review see
(Lovejoy and Schertzer, 2007). The purpose of this paper is to argue that the Haar
wavelet is particularly convenient, combining ease of calculation and of interpretation
and - if needed - it can easily be generalized to accommodate processes with H outside
this range. We demonstrate this using multifractal simulations with various H exponents
to systematically compare it to conventional spectral analysis and to another common –
and quite similar - analysis technique, the Detrended Fluctuation Analysis (DFA) method
(Peng et al., 1994), and its extension, the MultiFractal Detrended Fluctuation Analysis
(MFDFA) method (Kantelhardt et al., 2001), (Kantelhardt et al., 2002).
7
Fig. 1: This shows samples of the simulations analysed in section 5. From bottom to top,
H = -7/10, -3/10, 3/10, 7/10. All have C1 = 0.1, α =1.8, the sections are each 213 points
long. One can clearly see the change in character of the series when H changes sign, at
the bottom, fluctuations tend to cancel out, fluctuations are “stable”, at the top, they tend
to reinforce each other, fluctuations are “unstable”.
2. Defining fluctuations using wavelets
Our goal is to find a convenient definition of fluctuations valid at least over the
range -1<H<1, but before doing so, it will be useful to first define “tendency fluctuations
Δv Δx( )( )tend . The first step is to remove the overall mean ( v x( ) ) of the series:
′v x( ) = v x( )− v x( ) and then take averages over lag Δx. For this purpose, we introduce
the operator T Δx defining the tendency fluctuation Δv Δx( )( )tend as:
Δv Δx( )( )tend =T Δxv =1Δx
′v ′x( )x≤ ′x ≤x+Δx∑ (4)
or, with the help of the summation operator S , equivalently by:
Δv Δx( )( )tend =
1Δx
δΔxS v ' ; S v ' = ′v ′x( )′x ≤x∑ (5)
we can also use the suggestive notation Δv Δx( )( )tend = Δv , see the schematic fig. 2.
When -1<H<0, Δv Δx( )( )tend has a straightforward interpretation in terms of the mean
tendency of the data to decrease with averaging. The tendency fluctuation is also easy to
implement: simply remove the overall mean and then take the mean over intervals Δx:
this is equivalent to taking the mean of the differences of the running sum.
It is now straightforward to define the Haar fluctuation by taking the second
differences of the mean:
Δv Δx( )( )Haar = H Δxv =2Δx
δΔx/22 S v = 2
Δxs x( ) + s x + Δx( )( )− 2s x + Δx / 2( )( )
= 2Δx
v ′x( )x+Δx/2< ′x <x+Δx
∑ − v ′x( )x< ′x <x+Δx/2∑⎡
⎣⎢⎤⎦⎥; s x( ) = S v
(6)
8
Where H Δx is the Haar operator. Note the use of the shorthand notation s x( ) = S v . As
discussed in section 3, this is still a valid wavelet (see fig. 3), however it is almost trivial
to calculate and it can be used for -1<H<1 (appendix A; note there is an extra factor of
Δx-1 with respect to the Haar wavelet coefficient; see below). The numerical factor (2) is
used so that the Haar fluctuation is close to the usual differences when 0<H<1, (below).
While this may still look a little complicated it is actually quite simple: in words, the
Haar fluctuation at lag Δx is simply twice the difference of the mean from x to x+Δx/2
and from x+Δx/2 to x+Δx.
From the definitions, it is easy to obtain the relation:
H Δx = 2T Δx/2δΔx/2 = 2δΔx/2T Δx/2 (7) This is useful for interpreting the Haar fluctuations in terms of differences and tendencies
(section 5.2). Since the difference operator removes any constants, the tendency operator
can be replaced by an averaging operator and is thus insensitive to the addition of
constants; eq. 7 therefore means that the Haar fluctuation is simply the difference of the
series that has been degraded in resolution by averaging.
If needed, the Haar fluctuations can easily be generalized to higher order n by using
nth order differences on the sum s:
H Δx
n( )v =n +1( )Δx
δΔx / n+1( )n+1 s
(8) We note that the nth order Haar fluctuation is valid over the range -1<H<n; due to the
n+1th order difference, it is insensitive to polynomials of order n in the sum and of order
n-1 in the original series, although as we see below, for analyzing scaling series this
insensitivity has no particular advantage. Note further that H Δx1( ) = H Δx and
H Δx0( ) ′v =T Δx ′v hence when applied to a series with zero mean, H Δx
0( ) =T Δx .
Consider for a moment the case n = 2 which defines the “quadratic Haar
fluctuation”:
Δv Δx( )( )Haarquad =3Δx
s x + Δx( ) − 3s x + Δx / 3( ) + 3s x − Δx / 3( ) − s x − Δx( )( ) (9)
9
this fluctuation is sensitive to structures of size Δx-1 – and hence useful - over the range
-1<H<2 and it “filters” out polynomials of order 1 (lines); this is thus essentially
equivalent to the quadratic MFDFA (Multi Fractal Detrended Fluctuation Analysis,
section 4) technique that we numerically investigate below.
Fig. 2: The top shows extracts of a multifractal temperature simulation with H = 0.5, at
one hour resolution (in the weather regime), the bottom is the same simulation but at 16
day resolution (in the low frequency weather regime, H = -0.4). One can clearly see
their differing characters corresponding to H>0, H<0 respectively (see also fig. 1). These
simulations are used to illustrate the two different types of structure function needed
when H>0 (the usual) and H<0, the “tendency” structure function. Whereas the usual
structure function yields typical differences in T over an interval Δt, so that the mean is of
no consequence the tendency structure function uses the average of the field with the
mean removed. In both cases the q =2 moment was chosen because it is directly related
to the spectrum. Taking the square root is useful since the result is then a direct measure
of the “typical” fluctuations in units of K.
10
3. Fluctuations and Wavelets
Let’s now put these different fluctuations into the framework of wavelet analysis
where one defines fluctuations with the help of a basic “mother wavelet” Ψ(x) and
performs the convolution:
Δv x,Δx( ) = v ′x( )∫ Ψ ′x − xΔx
⎛⎝⎜
⎞⎠⎟d ′x
(10)
where we have kept the notation Δv to indicate “fluctuation”. Δv defined this way is
called a “wavelet coefficient”. The basic “admissibility” condition on Ψ(x) (so that it is a
valid wavelet) is that it has zero mean. If we take the fluctuation Δv as the symmetric
(centred) difference Δv = v x + Δx / 2( ) − v x − Δx / 2( ) this is equivalent to using:
Ψ x( ) = δ x −1 / 2( ) − δ x +1 / 2( ) (11)
where “δ” is the Dirac delta function (eq. 1, see fig. 3; due to the statistical translational
invariance, the coordinate shift by ½ is immaterial). As discussed, this “poor man’s”
wavelet is adequate for many purposes. As the generality of the definition (eq. 10)
suggests, all kinds of special wavelets can be introduced; for example orthogonal
wavelets can be used which are convenient if one wishes to reconstruct the original
function from the fluctuations, or to define power spectra locally in x rather than globally,
(averaged over all x).
The other simple to interpret fluctuation, the “tendency fluctuation” (eq. 4) can
also be expressed in terms of wavelets:
Ψ x( ) = I −1/2,1/2[ ] x( ) −I −L /2,L /2[ ] x( )
L; L >> 1
(12)
where I is the indicator function:
I a,b[ ] x( ) = 1 a ≤ x ≤ b0 otherwise (13)
11
see fig. 3 for a schematic. The first term in eq. 12 represents the sum whereas the second
removes the mean where L>>1 is the overall length of the data set. The removal of the
mean in this way is necessary so that the wavelet satisfy the admissibility condition that
its mean is zero. The only essential difference between the wavelet defined by eq. 12 and
the “tendency fluctuation” defined in eq. 4 (where the mean is removed beforehand), is
the extra normalization by Δx in eq. 4 which changes the exponent by one.
Figure 3 also shows the wavelet corresponding to the Haar fluctuation:
Ψ x( ) =1; 0 ≤ x <1/ 2
−1; −1/ 2 ≤ x < 0
0; otherwise
(14)
To within the division by Δx, and a factor of 2, the result is the equivalent of using the
Haar fluctuations (eq. 6) as indicated in fig. 3. It can be checked (appendix A) that the
Fourier transform of this wavelet is ∝ k−1 sin2 k / 4( ) so that compared to the difference
and tendency flucutations whose mother wavelet has cutoffs only at high or low
wavenumbers, it is better localized in Fourier space with both low and high wavenumber
fall-offs (≈ k and ≈ k-1 respectively). The Haar fluctuation is a special case of the
Daubechies family of wavelets (see e.g. (Holschneider, 1995) and for some applications,
see (Koscielny-Bunde et al., 1998; Koscielny-Bunde et al., 2006), (Ashok et al.,
2010)).
If we are interested in fluctuations valid up until H = 2, we need merely take
derivatives. For example, the second finite difference wavelet is:
Ψ x( ) = 12
δ x +12
⎛⎝⎜
⎞⎠⎟+ δ x −
12
⎛⎝⎜
⎞⎠⎟
⎛⎝⎜
⎞⎠⎟− δ x( )
(15)
(corresponding to Δv = (v(x+Δx)+v(x-Δx/2))/2 - v(x) using centred differences, fig. 4).
Alternatively, we can compare this with the quadratic Haar wavelet, eq. 9:
12
−1/ 3; −1< x < −1/ 3ψ x( ) = 2 / 3; −1/ 3< x <1/ 3
1/ 3;0;
1 / 3< x <1otherwise
(16)
which is very close to the Mexican hat wavelet (see fig. 4) but which is numerically much
easier to implement. As with the Mexican hat, it is valid for -1<H<2. It turns out that for
analyzing multifractals, that even over their common range of validity (0<H<1), the Haar
structure function is somewhat better than the poor man’s wavelet based structure
function (i.e. the usual structure function). This is demonstrated on numerical examples
in section 5. Just as one can use higher and higher order finite differences to extend the
range of H values, so one can simply use higher and higher order derivatives of the
Mexican hat.
13
Fig. 3: This shows the “poor man’s wavelet” by black bars representing the amplitudes
of Dirac δ functions, (the basis of the usual difference structure function, valid for 0<H<1,
this is only symbolic since the true δ function amplitude is of course infinite), the Haar
wavelet (the basis of the Haar structure function, uniform blue shading, the second
difference of the running sum, valid for -1<H<1), and the wavelet used for the “tendency”
structure function valid for -1<H<0), stippled shading.
Fig. 4: The popular Mexican hat wavelet (the second derivative of the Gaussian red,
valid for -1<H<2) compared with the (negative) second finite difference wavelet (black
bars representing the relative weights of δ functions, valid for 0<H<2)), and the second
order “quadratic” Haar wavelet (blue) eq. 9.
Wavelet mathematics are seductive - and there certainly exist areas such as speech
recognition - where both wavenumber and spatial localization of statistics are necessary.
However, the use of wavelets in geophysics is often justified by the existence of strong
localized structures that are cited as evidence that the underlying process is statistically
nonstationary/ statistically inhomogeneous (i.e. in time or in space respectively).
However, multifractal cascades produce exactly such structures - the “ singularities” - but
14
their statistics are strictly translationally invariant. In this case the structures are simply
the result of the strong singularities but nevertheless, the local statistics are usually
uninteresting, and we usually average over them in order to improve statistical estimates.
Before continuing to discuss other related methods for defining fluctuations, we
should mention that there have been strong claims that wavelets are indispensable for
analyzing multifractals, e.g. (Arneodo et al., 1999). In as much as the classical
definition of fluctuations as first differences is already a wavelet, this is true (see however
the DFA discussion below). However, for most applications, one doesn’t need the full
arsenal of wavelet techniques. At a more fundamental level however, the claim is at least
debatable since mathematically, wavelet analysis is a species of functional analysis, i.e.
its objects are mathematical functions defined at mathematical points. On the contrary,
the generic multifractal process – cascades - are rather densities of singular measures so
that strictly speaking, they are outside the scope of wavelet analysis. Although in
practice the multifractals are always cut-off at an inner scale, when used as a theoretical,
mathematical tool for multifractal analysis, wavelets may therefore not always be
appropriate. A particular example where wavelets may be misleading is the popular
“modulus maximum” technique where one attempts to localize the singularities (Bacry et
al., 1989), (Mallat and Hwang, 1992). If the method is applied to a cascade process then
as one increases the resolution, the singularity localization never converges implying that
the significance of the method is not as obvious as is claimed.
4: Fluctuations as deviations from polynomial regressions:
the DFA, MFDFA techniques
We now discuss a variant method that defines the fluctuations in a different way,
but still quantifies the statistics in the manner of the generalized structure function by
assuming that the fluctuations are stationary; “MultiFractal Detrended Fluctuation
Analysis”. The MFDFA technique (Kantelhardt et al., 2001), (Kantelhardt et al., 2002)
is a straightforward generalization of the original “Detrended Fluctuation Analysis”
(DFA, (Peng et al., 1994)). The method works for 1-D sections, so consider the
15
transect v(x) of series on a regular grid with resolution = 1 unit, N units long. The
MFDFA starts by replacing the original series by the running sum:
s x( ) = S v ' = ′v ′x( )
′x ≤x∑ (17)
(c.f. eq. 5). As discussed above, a sum is a finite difference of order -1 so that analysing s
rather than v allows the treatment of transects with H down to -1 (we will discuss the
upper bound shortly). We now divide the range into NΔx = int(N/Δx) disjoint intervals,
each indexed by i = 1, 2, …, NΔx (“int” means “integer part”). For each interval starting at
x = iΔx, one defines the “nth order” fluctuation by the standard deviation σs of the
difference of s with respect to a polynomial Fn(x = iΔx,Δx) = σs in the running sum of v
as follows:
Fn x = iΔx,Δx( ) = σ s =1Δx
s i −1( )Δx + j( )− pn,i j( )( )2j=1
Δx
∑⎡
⎣⎢
⎤
⎦⎥
1/2
; Δv Δx( )( )MFDFA =FnΔx
= σ s
Δx
(18)
where pn,ν(x) is the nth order polynomial regression of s(x) over the ith interval of length Δx.
It is the polynomial that “detrends” the running sum s; the fluctuation σs is the root mean
square deviation from the regression. F is the standard DFA notation for the “fluctuation
function”, as indicated, it is in fact a fluctuation in the integrated series quantified by the
standard deviation σs. Using F we have Fn= σs = (Δv)MFDFAΔx so in terms of the usually
cited DFA fluctuation exponent α (not the Levy α!) we have Fn ≈ Δxα and α =1+H so that
the (second order) spectral exponent β = 2α-1-K(2), so that the oft-cited relation β = 2α-1
ignores the intermittency correction K(2). The fluctuation (Δv) MFDFA is very close to the
nth order finite difference of v, it is also close to a wavelet based fluctuation where the
wavelet is of nth order, although this is not completely rigorous (see however
(Kantelhardt et al., 2001; Taqqu et al., 1995)). As a practical matter, in the above sum,
we start with intervals of length Δx = n+1 and normalize by the number of degrees of
freedom of the regression, i.e. by Δx-n not by Δx; formula 18 is the approximation for
large Δx.
16
In the usual presentation of the MFDFA, one considers only a single realization of
the process and averages powers of the fluctuations over all the disjoint intervals i.
However, more generally, we can consider a statistical ensemble and average over all the
intervals and realizations. If in addition, we consider moments other than q = 2, then we
obtain the MultiFractal DFA (MFDFA) with the following statistics:
σ s Δx( ) q1/q
= Δx Δv Δx( )( )MFDFAq 1/q
= Δxh q( ); h(q) = 1+ ξ(q) / q (19)
where we have used the exponent h(q) as defined in (Davis et al., 1996) and
(Kantelhardt et al., 2002) and the relation Δv( )MFDFAq ≈ Δxξ q( ) . In what follows, we will
refer to the method as the MFDFA technique even though the only difference with
respect to the DFA is the consideration of the q ≠ 2 statistics.
From our analysis above, we see that the nth order MFDFA exponent h(q) is
related to the usual exponents by:
h q( ) = 1+ ξ q( ) / q = 1+ H( )− K q( ) / q; −1< H < n (20)
Note that in eq. 20, the 1 appears because of the initial integration of order 1 in the
MFDFA recipe, the method works up to H = n because the fluctuations are defined as the
residues with respect to nth order polynomials of the sum s corresponding to n+1th order
differences in s and hence n-1th order differences in v). The justification for defining the
MFDFA exponent h(q) in this way is that in the absence of intermittency (i.e. if K(q) =
constant = K(0) = 0), one obtains h(q) = H = constant. Unfortunately - contrary to what is
often claimed - the nonconstancy of h(q) - neither implies that that the process is
multifractal nor conversely does its constancy imply a monofractal process. To see this,
it suffices to consider the (possibly fractionally integrated, order H) monofractal β model
which has K(q) = C1(q-1) so that h(q) = 1+H- C1+C1/q which is not only non constant but
even diverges as q approaches zero (indeed the same q = 0 divergence occurs for any
universal multifractal with α<1, see eq. 21 below).
During the last ten years, the MFDFA technique has been applied frequently to
atmospheric data, of special note is the work on climate by A. Bunde and co-workers (e.g.
17
(Koscielny-Bunde et al., 1998), (Kantelhardt et al., 2001), (Bunde et al., 2002), (Bunde
et al., 2005), (Lennartz and Bunde, 2009)). However, the exact status of the MFDFA
method is not clear since a completely rigorous mathematical analysis has not been done
(especially in the multifractal case), and the literature suffers from misleading claims to
the effect that the MFDFA is necessary because by its construction it removes
nonstationarities due to trends (linear, quadratic etc. up to polynomials order n-1) in the
data. While it is true that it does remove polynomial trends, these only account for fairly
trivial types of nonstationarity: beyond this, the method continues to make the standard
(and strong) stationarity assumptions about the statistics of the deviations which are left
over after performing this polynomial detrending. In particular it does nothing to remove
the most common genuine type of statistical nonlinearity in the geosciences: the diurnal
and annual cycles which still strongly break the scaling of the MFDFA statistics, and this
for any order n. Furthermore, the corresponding wavelet or finite difference definition of
fluctuations can equally well take into account these polynomial trends, and finally, the
method removes trends at all scales and locations so that this emphasis on detrending is
misleading. In other words, the MFDFA is a variant with respect to some of the wavelet
methods, in particular with respect to the Haar wavelet and its generalizations. While it
may indeed yield accurate exponent estimates it has the disadvantage that the
interpretation of the fluctuations is not straightforward. In comparison, the usual
difference and tendency fluctuations and structure functions have simple physical
interpretations in terms of magnitudes of changes (when 1>H>0) and magnitudes of
average tendencies (when -1<H<0).
5. A comparison of different fluctuations using multifractal
simulations
5.1 The simulations:
To have a clearer idea of the limitations of the various fluctuations for determining
the statistics of scaling functions, we shall numerically compare their performance and
their corresponding structure functions when applied to the characterization of
18
multifractals. In order to easily compare them with standard spectral analysis, we will
only consider the second order (q = 2) structure functions and will use simulations of
universal multifractals. Recall that universal multifractals are the result of stable,
attractive multiplicative cascade processes and have:
K q( ) = C1α −1
qα − q( ) (21)
where 0≤C1 is the codimension of the mean and 0≤α≤2 is the multifractality index (the
Lévy index of the generator). The limit α->1 gives K(q) = C1qlogq and may be obtained
from eq. 21 by using l’Hopital’s rule.
The following tests were all made on simulations with parameters α=1.8, C1 = 0.1
with H in the range -7/10<H<7/10 (typical parameters in meteorology and in solid earth
geophysics, see the reviews (Lovejoy and Schertzer, 2010b), (Lovejoy and Schertzer,
2007)). These parameters yield an intermittency correction K(2) = 0.18; 50 simulations
were averaged to estimate the ensemble statistics. The simulations were made using the
technique described in (Schertzer and Lovejoy, 1987) and refined in (Lovejoy and
Schertzer, 2010a) for the flux ϕ (i.e. H = 0). For the second fractional integration (to
obtain the field v from ϕ), we used a Fourier space power law filter k-H. For this
fractional integration, exponential cutoffs were used at high wavenbumbers to avoid
small scale numerical instabilities (this is standard for differentiation, i.e. when filtering
by k-H with H <0, it is equivalent to using the “Paul” wavelet; see (Torrence and Compo,
1998)). The raw series were 216 in length and were then degraded in resolution by
factors of 4 to 214 to avoid residual finite size effects in the simulation at the smallest
scales. In order to avoid spurious correlations introduced by the periodicity of the
simulations, the series were split in half so that the largest scale was 213.
In fig. 1 we already showed samples of the simulations visually showing the effect
of increasing H from -7/10 to 7/10. In fig. 5, we show the compensated spectra (obtained
using standard Hann windows to avoid spectral leakage). As expected, we see that they
are nearly flat, although for the lowest and highest factors of about 100.5 ≈3 there were
more significant deviations; the slopes of the central portions were thus determined by
regression and the corresponding exponents are shown in fig. 6. In fig. 5, the spectra
19
were averaged over logarithmically spaced bins, 10 per order of magnitude (with the
exception of the lowest factor of 10 where all the values were used). We also performed
the regressions using log-log fits using all the Fourier components (rather than just the
averaged values) in the same central range, these gave virtually identical exponent
estimates (the mean absolute differences in the estimated spectral exponent β were ≈
±0.007), and were thus not shown. Similarly, if instead of using log-log linear regression,
we use nonlinear regressions over the same range we obtain identical exponents to within
±0.009, and again these estimates were too similar to be worth showing. From fig. 6,
one can see that there seems to be a residual small bias of about -0.04 the origin of which
is not clear.
Fig. 5: The compensated spectra for an ensemble of 50 realizations, 214 each, α = 1.8, C1
= 0.1, intermittency correction = K(2) = 0.18, with H increasing from top to bottom from
-7/10 to 7/10. The dashed horizontal line is the theoretical behaviour indicated over the
range used to estimate the exponent (i.e. the highest and lowest factor of 100.5 in
wavenumber has been dropped). Each curve was offset in the vertical for clarity.
20
Fig. 6: This shows the regression estimates of the compensated exponents for the spectra
(red), Haar structure function (q =2, green), quadratic, q =2 MFDFA (dashed), The usual
difference (poor man’s) structure function (q =2, blue, for H>0), and the tendency
structure function (q =2, same line, blue for H<0).
21
Fig. 7: This shows the compensated Haar structure functions
SHaar Δx( ) = ΔvHaar Δx( )21/2
(solid lines) and the second order MFDFA technique
(dashed lines), again for H = -7/10 (top) to H = 7/10 (bottom), every 1/5. Each curve was
offset in the vertical for clarity by ¼.
22
Fig. 8: This compares the compensated Haar structure function (thick), the difference
structure function (thin, below the axis, for H =3/10 (bottom) and 1/10 orange, second
up), and the tendency structure function (third from bottom, thin red, H = -1/10), and top,
thin green, H = -3/10. It can be seen that the standard difference structure function has
poor scaling for nearly two orders of magnitude when H =1/10, and one order of
magnitude for H =3/10 (see fig. 7 for quantitative estimates).
We next consider the Haar structure function which we compare to the second
order, quadratic MFDFA structure function analogue (based on the MFDFA fluctuation
(Δv)MFDFA = F/Δx where F is the usual MFDFA scaling function, see eq. 18), the former is
valid over the range -1<H<1, the latter over the range -1<H<2. We can see (fig. 7) that
the Haar structure function does an excellent job with overall bias mostly around +0.01 to
0.02, (fig. 6), and that the MFDFA method is nearly as good (overall bias ≈ -0.02 to -
0.04). It is interesting to compare this in the same figure with the results of the tendency
structure function (H<0) and the usual difference (poor man’s) structure function (H>0),
fig. 8. The tendency structure function turns out to be quite accurate, except near the
limiting value H = 0 with the large scales showing the largest deviations. In comparison,
for the difference structure function, the scaling is poorest at the smaller scales requiring
a range factor of ≈ 100 convergence for H = 1/10 and a factor ≈ 10 for H = 3/10. The
23
overall regression estimates (fig. 6) show that the biases for both increase near their
limiting values H = 0 to accuracies ≈ ±0.1 (note that they remain quite accurate if one
only makes regressions around the scaling part). This can be understood since the
process can be thought of as a statistical superposition of singularities of various orders
and only those singularities with large enough orders (difference structure functions) or
small enough orders (tendency structure functions) will not be affected by the theoretical
limit H = 0. The main advantage of the Haar structure function with respect to the usual
(tendency or difference) structure functions is precisely that it is valid over the whole
range -1< H <1 so that it is not sensitive to the H =0 limit.
Before proceeding, we could make a general practical remark about these real space
statistics: most lags Δx do not divide the length of the series (L) exactly so that there is a
“remainder” part. This is not important for the small Δx<<L, for each realization there
are many disjoint intervals of length Δx however when L/3≈<Δx<L the statistics can be
sensitive to this since from each realization there will only one or two segments and
hence poor statistics. A simple expedient is to simply repeat the analysis on the reversed
series and average the two results. This can indeed improve the statistics at large Δx
when only one or a small number of realizations are available; it has been done in the
analyses below.
5.2 The theoretical relation between poor man’s, tendency and
Haar fluctuations, hybrid structure functions
We have seen that although the difference and tendency structure functions have
the advantage of having simple interpretations in terms of respectively the average
changes in the value of the process and its mean value over an interval, this simple
interpretation is only valid over a limited range of H values. In comparison, the Haar and
MFDFA fluctuations give structure functions with valid scaling exponents over wider
ranges of H. But what about their interpretations? We now briefly show how to relate
the Haar, tendency and poor man’s structure functions thus giving it a simple
interpretation as well.
24
In order to see the connection between the fluctuations, we use the “saturation”
relations:
δΔxv=dCtendv; H < 0
T Δxv=dCdiff v; H > 0
(22)
where Ctend, Cdiff are proportionality constants and =d
indicates equality in the random
variables in the sense of probability distributions ( a=db if Pr(a>ζ) = Pr(b> ζ) where “Pr”
means “probability” and ζ is an arbitrary threshold). These relations were easily verified
numerically and arise for the reasons stated above; they simply mean that for series with
H<0, the differences are typically of the same order as the function itself and that for H>0,
the tendencies are of the same order: the fluctuations “saturate”. By using eqs. 7 and eq.
22 we now obtain:
H Δxv = 2T Δx /2δΔx /2v=d2T Δx /2v=
d′CtendT Δxv; H < 0
H Δxv = 2δΔx /2T Δx /2v=d2δΔx /2v=
d′CdiffδΔxv; H > 0
(23) where C’tend, C’diff are “calibration” constants (only a little different from the unprimed
quantities – they take into account the factor of two and the change from Δx/2 to Δx).
Taking the qth moments of both sides of eq. 23, we obtain the results for the various qth
order structure functions:
Δv( )Haarq
Δv( )diffq = ′Cdiff
q ; H > 0
Δv( )Haarq
Δv( )tendq = ′Ctend
q ; H < 0
(24) This shows that at least for scaling processes that the Haar structure functions will be the
same as the difference (H>0) and tendency structure functions (H<0), as long as these are
“calibrated” by determining C’diff, and C’tend. In other words, by comparing the Haar
structure functions with the usual structure functions we can develop useful correction
factors which will enable us to deduce the usual fluctuations given the Haar fluctuations.
25
Although the MFDFA fluctuations are not wavelets, at least for scaling processes, the
same basic argument applies, since the MFDFA and usual fluctuations scale with the
same exponents, they can only differ in their prefactors, the calibration constants.
To see how this works on a scaling process, we confined ourselves to considering
the root mean square (RMS) S(Δt) = <Δv2>1/2; see fig. 9 which show the ratio of the usual
S(Δt) to those of the RMS Haar and MFDFA fluctuations. We see that at small Δx’s, the
ratio stabilizes after a range of about a factor 10 – 20 in scale and that it is quite constant
up to the extreme factor of 3 or so (depending a little on the H value). The deviations are
largely due to the slower convergence of the usual RMS fluctuations to their asymptotic
scaling form. From the figure, we can see that the ratios C’Haar = S/SHaar (fig. 9 upper left)
and corresponding C’MFDFA = S/SMFDFA (fig. 9, upper right) are well defined in the central
region, and are near unity, more precisely in the range ½< C’Haar <2 for the most
commonly encountered range of H: -4/10<H<4/10. In comparison (fig. 9, bottom) over
the same range of H, we have 0.23> C’MFDFA >0.025 so that the MFDFA fluctuation is
quite far from the usual ones and varies strongly with H. For applications, one may use
the semi-empirical formulae C’Haar ≈ 1.1 e-1.65H and C’MFDFA ≈ 0.075e-2.75H which are quite
accurate over the entire range -7/10<H<7/10. Although these factors in principle allow
one to deduce the usual RMS structure function statistics from the MFDFA and Haar
structure functions this is only true in the scaling regime. Since the corrections for Haar
fluctuations are close to unity for many purposes they can be used directly.
26
Fig. 9: These graphs compare the ratios of RMS structure functions ( S Δx( ) = Δv21/2
)
for the simulations discussed in the text.
Upper left: This shows the ratio of the RMS usual structure function (i.e. difference
when H>0, tendency when H<0) to the Haar structure function for H =1/10, 3/10,.. 9/10
(thick, top to bottom), and -9/10, -7/10, … -1/10, (thin, top to bottom). The Haar
structure function has been multiplied by 2ξ(2)/2 to account for the difference in effective
resolution. The central flat region is where the scaling is accurate indicating a constant
ratio C’Haar = S/SHaar which is typically less than a factor of two.
Upper right: The same but for the ratio of the usual to MFDFA RMS structure functions
after the MFDFA was normalized by a factor of 16 and the resolution correction 4ξ(2)/2 (its
smallest scale is 4 pixels).
Bottom left: This shows the ratio of the MFDFA structure function to the Haar structure
function both with the normalization and resolution corrections indicated above (note the
monotonic ordering with respect to H which increases from top to bottom).
27
5.3: Hybrid fluctuations and structure functions: an empirical
example with both H<0, and H>0 regimes
For pure scaling functions, the difference (1>H>0) or tendency (-1<H<0) structure
functions are adequate, the real advantage of the Haar structure function is apparent for
functions with two or more scaling regimes one with H>0, one with H<0. Can we still
“calibrate” the Haar structure function so that the amplitude of typical fluctuations can
still be easily interpreted? For these, consider eqs. 23 which motivate the definition:
H hybrid ,Δxv = max δΔxv,T Δxv( )
(25) of a “hybrid” fluctuation as the maximum of the difference and tendency fluctuations; the
“hybrid structure function” is thus the maximum of the corresponding difference and
tendency structure functions and therefore has a straightforward interpretation. The
hybrid fluctuation is useful if a unique calibration constant Chybrid can be found such that:
H hybrid ,Δxv≈dChybridH Δxv (26)
To get an idea of how the different methods can work on real data, we consider
the example of the global monthly averaged surface temperature of the earth. This is
obviously highly significant for the climate but the instrumental temperature estimates
are only known over a small number of years.
For this purpose, we chose three different series whose period of overlap was
1881-2008; 129 years. These were the NOAA NCDC (National Climatic Data Center) merged land air and sea surface temperature dataset (abbreviated NOAA NCDC below,
from 1880 on a 5°x5° grid, see (Smith et al., 2008) for details), the NASA GISS
(Goddard Institute for Space Studies) data set (from 1880 on a 2°x2° (Hansen et al.,
2010) and the HadCRUT3 data set (from 1850 to 2010 on a 5° x5° grid). HadCRUT3 is
a merged product created out of the HadSST2 (Rayner et al., 2006) Sea Surface
Temperature (SST) data set and its companion data set CRUTEM3 of atmospheric
temperatures over land, see (Brohan et al., 2006). Both the NOAA NCDC and the
NASA GISS data were taken from http://www.esrl.noaa.gov/psd/; the others from
http://www.cru.uea.ac.uk/cru/data/temperature/. The NOAA NCDC and NASA GISS are
both heavily based on the Global Historical Climatology Network (Peterson and Vose,
28
1997), and have many similarities including the use of sophisticated statistical methods
to smooth and reduce noise. In contrast, the HadCRUTM3 data is less processed with
corresponding advantages and disadvantages.
Fig. 10 shows the comparison of the difference, tendency, hybrid and Haar RMS
structure functions; the latter increased by a factor Chybrid = 100.35 ≈ 2.2. It can be seen that
the hybrid structure function does extremely well; the deviation of the calibrated Haar
structure function from the hybrid one is ±14% over the entire range of near a factor 103
in time scale. This shows that to a good approximation, the Haar structure function can
preserve the simple interpretation of the difference and tendency structure functions: in
regions where the logarithmic slope is between -1 and 0, it approximates the tendency
structure function whereas in regions where the logarithmic slope is between 0 and 1, the
calibrated Haar structure function approximates the difference structure function.
Before embracing the Haar structure function let us consider it’s behaviour in the
presence of nonscaling perturbations; it is common to consider the sensitivity of
statistical scaling analyses to the presence of nonscaling external trends superposed on
the data which therefore break the overall scaling. Even when there is no reason to
suspect such trends, the desire to filter them out is commonly invoked to justify the use of
quadratic MFDFA or high order wavelet techniques that eliminate linear or higher order
polynomial trends. However, for this purpose, these techniques are not obviously
appropriate since on the one hand they only filter out polynomial trends (and not for
example the more geophysically relevant periodic trends), while on the other hand, even
for this, they are “overkill” since the trends they filter are filtered at all – not just the
largest scales. The drawback is that with these higher order fluctuations, we loose the
simplicity of interpretation of the Haar fluctuation while obtaining few advantages. Fig.
11 shows the usual (linear) Haar RMS structure function compared to the quadratic Haar
and quadratic MFDFA structure functions. It can be seen that while the latter two are
close to each other (after applying different calibration constants, see the figure caption),
and that the low and high frequency exponents are roughly the same, the transition point
has shifted by a factor of about 3 so that overall they are quite different from the Haar
structure function, it is therefore not possible to simultaneously calibrate the high and low
frequencies.
29
Fig. 10: A comparison of the different structure function analyses (root mean
square, RMS) applied to the ensemble of three monthly surface series discussed in
appendix 10 C (NASA GISS, NOAA CDC, HADCRUT3), each globally and annually
averaged, from 1881-2008 (1548 points each). The usual (difference, poor man’s)
structure function is shown (orange, lower left), the tendency structure function (purple,
lower right), the maximum of the two (“Hybrid”, thick, red), and the the Haar in blue (as
indicated); it has been increased by a factor C =100.35 = 2.2 and the RMS deviation with
respect to the hybrid is ±14%. Reference slopes with exponents ξ(2)/2 ≈ 0.4, -0.1 are
also shown (black corresponding to spectral exponents β = 1+ξ(2) = 1.8, 0.8 respectively).
In terms of difference fluctuations, we can use the global root mean square ΔT Δt( )21/2
annual structure functions (fitted for 129 yrs > Δt >10 yrs), obtaining ΔT Δt( )21/2
≈
0.08Δt 0.33 for the ensemble, in comparison, (Lovejoy and Schertzer, 1986) found the very
similar ΔT Δt( )21/2
≈ 0.077Δt0.4 using northern hemisphere data (these correspond to βc =
1.66, 1.8 respectively). Reproduced from (Lovejoy and Schertzer, 2012a).
30
Fig. 11: The same temperature data as fig. 10; a comparison of the RMS Haar structure
function (multiplied by 100.35 = 2.2), the RMS quadratic Haar (multiplied by 100.15 = 1.4)
and the RMS quadratic MFDFA (multiplied by 101.5 = 31.6). Reproduced from (Lovejoy
and Schertzer, 2012a).
If there are indeed external trends that perturb the scaling, these will only exist at
the largest scales; it is sufficient to remove a straight line (or if needed second, third or
higher order polynomials) from the entire data set i.e. at the largest scales only. Fig. 12
shows the result when the Haar structure function is applied to the data that has been
detrended in two slightly different ways: by removing a straight line through the first and
last points of the series and by removing a regression line. Since these changes
essentially effect the low frequencies only, they mostly affect the extreme factor of two in
scale. If we discount this extreme factor 2, then in the figure we see that there are again
two scaling regimes, the low frequency one is slightly displaced with respect to the Haar
structure function, but not nearly as much as for the quadratic Haar. In other words,
removing the linear trends at all scales as in the quadratic Haar or the quadratic MFDFA
is too strong to allow a simple interpretation of the result and it is unnecessary if one
wishes to eliminate external trends.
31
Fig. 12: The same temperature data as fig. 10; a comparison of the RMS Haar structure
function applied to the raw data, after removing linear trends in two ways. The first
defines the linear trend by the first and last points while the second uses a linear
regression; all were multiplied by Chybrid = 100.35 = 2.2.
In summary, if all that is required are the scaling exponents, the Haar structure
function and quadratic MFDFA seem to be a excellent techniques. However, the Haar
structure function has the advantage of numerical efficiency, simplicity of
implementation, simplicity of interpretation.
6. Conclusions:
Geosystems are commonly scaling over large ranges of scale in both space and time.
In order to characterize their properties, one decomposes them into elementary
components with well defined scales and characterizes the variability with the help of
various statistics. Their scaling behaviours can then be characterized by exponents
obtained by systematically changing scale. The classical analysis techniques are spectral
analysis (Fourier decomposition) and structure functions based fluctuations defined as
differences. However, over the last twenty years, the development of wavelets has
provided a systematic mathematical and fairly general basis for defining fluctuations. As
32
far as geophysical applications are concerned, this generality has not proved to be so
helpful. This is because different workers use different wavelets and these are often
chosen for various mathematical properties which are not necessarily important for
applications. In addition, the numerical determination of the wavelet coefficients is often
cumbersome. Finally, the interpretation of the corresponding wavelet coefficients is
often obscure and attention has been largely focused on the exponents.
These features of wavelet analysis have lead to the continued popularity of a
slightly different wavelet- like technique - the Detrended Fluctuation Analysis (DFA)
method and its extension, the MultiFractal DFA (MFDFA). An unfortunate consequence
is that the criterion for the applicability of the different techniques has largely been
ignored (the range of H exponents), and the advantages and disadvantages of different
approaches has not been critically (and quantitatively) evaluated. In this paper, we argue
that the historically first wavelet - the Haar wavelet - has several advantages. These are
not so much at the level of the accuracy of the exponents – which turn out to be about the
same as for spectral analysis and the DFA technique – but rather in the ease of
implementation and interpretation of the results. This is demonstrated both with the help
of multifractal simulations (exponents) but also the ease of interpretation with an
atmospheric example where we show that with “calibration” the Haar structure functions
can be made very close to both the usual difference structure functions (in regions where
1>H>0) and to simple “tendency” structure functions (in regions where -1<H<0). This
simplicity is already lost when using either the DFA or even higher order wavelet based
generalizations of Haar fluctuations. The Haar structure functions have recently been
applied systematically to atmospheric data where they have clearly shown that there are
not two regimes (weather and climate) but rather three (weather <≈10 days, low
frequency weather (between ≈10 days and ≈ 10-30 years and the climate ≈>10 - 30 years),
(Lovejoy and Schertzer, 2012a).
33
Appendix A: Relation between real and Fourier space
analyses and the limiting H exponents
We repeatedly underlined the limiting H values associated with specific type of
fluctuations. In this appendix we clarify the origin of these limits by considering the
relation between the fluctuations and the spectra.
Consider the statistically translationally invariant process v(x) in 1-D: the statistics
are thus independent of x and this implies that the Fourier components are “δ correlated”:
v k( )v ′k( ) = δk+ ′k v k( ) 2 ; v k( ) = e− ikxv x( )dx∫
(27)
If it is also scaling then the spectrum E(k) is a power law: E k( ) ≈ v k( ) 2 ≈ k−β
(where
here and below, we ignore constant terms such as factors of 2π etc.). Let us first consider
the difference fluctuation. In terms of its Fourier components, this fluctuation is thus:
Δv x,Δx( )( )diff = v x + Δx( )− v x( ) = eikx v k( ) eikΔx −1( )dk∫ (28)
so that the F.T. of (Δv(x,Δx))diff is v k( ) eikΔx −1( ) . We first consider the statistics of quasi-
Gaussian processes for which C1 = 0, ξ(q) = Hq. Assuming statistical translational
invariance, we drop the x dependence and relate the second order structure function to the
spectrum:
Δv Δx( ) 2 = 4 eikx v k( )2 sin2 kΔx
2⎛⎝⎜
⎞⎠⎟dk∫ ≈ eikxk−β sin2 kΔx
2⎛⎝⎜
⎞⎠⎟dk∫
(29)
As long as the integral on the right converges, then using the transformation of variables,
x→λx , k→λ−1k , we find:
Δv Δx( ) 2 ∝ Δxξ 2( ) ≈ eikxk−β sin2 kΔx2
⎛⎝⎜
⎞⎠⎟dk∫ ∝ Δxβ−1
(30)
34
so that β = ξ(2)+1 = 2H +1 (C1 = 0 here). However, for large k, the integrand ≈ k-β which
has a large wavenumber divergence whenever β<1. However, since for small k,
sin2 kΔx / 2( )∝ k2 , there will be a low wavenumber divergence only when β>3. In the
divergent cases, although real world (finite) data will not diverge, the structure functions
will no longer characterize the fluctuations, but will depend spuriously on the either the
highest or lowest wavenumbers present in the data. In the quasi Gaussian case, or when
C1 is small, we have ξ(2) ≈ 2H and we conclude that the using first order differences to
define the fluctuations leads to second order structure functions being meaningful in the
sense that they adequately characterize the fluctuations whenever 1<β<3, i.e. 0<H<1.
Since 0<H<1 is the usual range of geophysical H values, and the difference
fluctuations are very simple, they are commonly used. However, we can see that there
are limitations; in order to extend the range of H values, it suffices to define fluctuations
using finite differences of different orders. To see how this works, consider using second
(centred) differences:
Δv x,Δx( )( )second = v x( )− 12v x + Δx( )− v x − Δx( )( ) = eikx v k( ) 1− 1
2eikΔx/2 + e−ikΔx/2( )⎡
⎣⎢⎤⎦⎥dk∫
= 2 eikx v k( )sin kΔx4
⎛⎝⎜
⎞⎠⎟2
dk∫ (31)
repeating the above arguments we can see that the relation β = ξ(2)+1 holds now for
1<β<5, or (with the same approximation) 0<H<2. Similarly, by replacing the original
series by its running sum (a finite difference of order -1) – as done in the MFDFA
technique – we can extend the range of H values down to -1.
More generally, going beyond Gaussian processes we can consider intermittent
Fractionally Integrated Flux processes (section 5.1), which have v k( ) ≈ ε k( ) k −H
, we see
that the F.T. of (Δv(x,Δx))diff is eikΔx −1( ) k −H ε k( ) . This implies that for low
wavenumbers, v k( ) ≈ k 1−H ε k( ) (k<<1/Δx) whereas at high wavenumbers,
v k( ) ≈ k −H ε k( )
(k>>1/Δx) hence since the second characteristic function of ε(x) has
35
logarithmic divergences with scale for both large and small scales ( log ελq = K q( ) logλ ),
we see that for 0<H<1, that the fluctuations Δv(Δx) are dominated by wavenumbers k ≈
1/Δx, so that for this range of H, fluctuations defined as differences capture the variability
of Δx sized structures, not structures either much smaller or much larger than Δx. More
generally, since the F.T. of the nth derivative dnv/dxn is ik( )n v k( ) and the finite derivative
is the same for small k but “cut-off” at large k, we find that nth order fluctuations are
dominated by structures with k ≈1/Δx as long as 0 < H <n. This means that Δv(Δx) does
indeed reflect the Δx scale fluctuations.
References
Arneodo, A., Decoster, N. and Roux, S.G., 1999. Intermittency, Log-Normal Statistics,
And Multifractal Cascade Process In High-Resolution Satellite Images Of Cloud
Structure, Physical Review Letters 83, pp. 1255-1258.
Ashok, V., Balakumaran, T., Gowrishankar, C., Vennila, I.L.A. and Nirmalkumar, A.,
2010. The Fast Haar Wavelet Transform for Signal & Image Processing, , Inter. J.
of Comp Sci, Info. Security 7, pp. 126-130.
Bacry, A., Arneodo, A., Frisch, U., Gagne, Y. and Hopfinger, E., 1989. Wavelet analysis
of fully developed turbulence data and measurement of scaling exponents. In: M.
Lessieur and O. Metais, Editors, Turbulence and coherent structures, Kluwer, pp.
703-718.
Brohan, P., Kennedy, J.J., Harris, I., Tett, S.F.B. and Jones, P.D., 2006. Uncertainty
estimates in regional and global observed temperature changes: a new dataset
from 1850, J. Geophysical Research 111, p. D12106.
Bunde, A., Eichner, J.F., Kantelhardt, J.W. and Havlin, S., 2005. Long-term memory: a
natural mechanism for the clustering of extreme events and anomalous residual
times in climate records, Phys. Rev. Lett. 94, p. 048701.
Bunde, A., Kropp, J. and Schellnhuber, H.J., Editors, The science of disasters: Climate
disruptions, heart attacks and market crashes Springer (2002).
36
Davis, A., Marshak, A., Wiscombe, W. and Cahalan, R., 1996. Multifractal
characterization of intermittency in nonstationary geophysical signals and fields.
In: J. Harding, B. Douglas and E. Andreas, Editors, Current topics in
nonstationary analysis, World Scientific, Singapore, pp. 97-158.
Haar, A., 1910. Zur Theorie des orthogonalen Funktionsysteme, Mathematische Annalen
69, pp. 331-371.
Hansen, J., Ruedy, R., Sato, M. and Lo, K., 2010. Global surface temperature change,
Rev. Geophys. 48, p. RG4004.
Holschneider, M., 1995. Wavelets : an analysis tool, Oxford : Clarendon Press, New York
Kantelhardt, J.W., Koscielny-Bunde, E., Rego, H.H.A., Havlin, S. and Bunde, S., 2001.
Detecting long range corelations with detrended flucutation analysis, Physica A
295, pp. 441-454.
Kantelhardt, J.W., Zscchegner, S.A., Koscielny-Bunde, K., Havlin, S., Bunde, A. and
Stanley, H.E., 2002. Multifractal detrended fluctuation analysis of nonstationary
time series, Physica A 316, pp. 87-114.
Koscielny-Bunde, E., Bunde, A., Havlin, S., Roman, H.E., Goldreich, Y. and
Schellnhuber, H.J., 1998. Indication of a universal persistence law governing
atmospheric variability, Phys. Rev. Lett. 81, p. 729.
Koscielny-Bunde, E., Kantelhardt, J.W., Braund, P., Bunde, A. and Havlin, S., 2006.
Long-term persistence and multifractality of river runoff records: Detrended
fluctuation studies, Journal of Hydrology 322, pp. 120–137.
Lennartz, S. and Bunde, A., 2009. Trend evaluation in records with long term memory:
Application to global warming, Geophys. Res. Lett. 36, p. L16706.
Lovejoy, S. and Schertzer, D., 1986. Scale invariance in climatological temperatures and
the spectral plateau, Annales Geophysicae 4B, pp. 401-410.
Lovejoy, S. and Schertzer, D., 2007. Scaling and multifractal fields in the solid earth and
topography, , Nonlin. Processes Geophys. 14 pp. 1-38.
Lovejoy, S. and Schertzer, D., 2010a. On the simulation of continuous in scale universal
multifractals, part I: spatially continuous processes, Computers and Geoscience, p.
10.1016/j.cageo.2010.1004.1010.
37
Lovejoy, S. and Schertzer, D., 2010b. Towards a new synthesis for atmospheric
dynamics: space-time cascades, Atmos. Res. 96, pp. pp. 1-52,
doi:10.1016/j.atmosres.2010.1001.1004.
Lovejoy, S. and Schertzer, D., 2012a. Low frequency weather and the emergence of the
Climate. In: A.S. Sharma, A. Bunde, D. Baker and V.P. Dimri, Editors,
Complexity and Extreme Events in Geosciences, AGU monographs.
Lovejoy, S. and Schertzer, D., 2012b. The Weather and Climate: Emergent Laws and
Multifractal Cascades, Cambridge University Press, Cambridge 660 pp.
Mallat, S. and Hwang, W., 1992. Singularity detection and processing with wavelets,
IEEE Trans. Info. Theory 38, pp. 617-643.
Peng, C.-K., Buldyrev, S.V., Havlin, S., Simons, M., Stanley, H.E. and Goldberger, A.L.,
1994. Mosaic organisation of DNA nucleotides, Phys. Rev. E 49, pp. 1685-1689.
Peterson, T.C. and Vose, R.S., 1997. An overview of the Global Historical Climatology
Network temperature database, Bull. Amer. Meteorol. Soc. 78, pp. 2837-2849.
Rayner, N.A., Brohan, P., Parker, D.E., Folland, C.K., Kennedy, J.J., Vanicek, M., Ansell,
T. and Tett, S.F.B., 2006. Improved analyses of changes and uncertainties in
marine temperature measured in situ since the mid-nineteenth century: the
HadSST2 dataset. , J. Climate 19, pp. 446-469.
Schertzer, D. and Lovejoy, S., 1987. Physical modeling and Analysis of Rain and Clouds
by Anisotropic Scaling of Multiplicative Processes, Journal of Geophysical
Research 92, pp. 9693-9714.
Smith, T.M., Reynolds, R.W., Peterson, T.C. and Lawrimore, J., 2008. Improvements to
NOAA's Historical Merged Land-Ocean Surface Temperature Analysis (1880-
2006), J. Climate 21, pp. 2283-2293.
Taqqu, M., Teverovsky, V. and Willinger, W., 1995. Fractals 3, p. 785.
Torrence, T. and Compo, G.P., 1998. A practical guide to wavelet analysis, Bull. of the
Amer. Meteor. Soc. 79, pp. 61-78.